Instate documentation¶
instate: state and language composition estimates for Indian surnames¶
Instate looks up electoral-roll surname shares across 35 states and union territories and estimates shares for unseen surnames with a calibrated character model. It derives language compositions by mixing state shares with Census 2011 mother-tongue shares.
The lookup contains 1,855,202 surname strings. Lookup and training retain only surname-state cells with at least three source occurrences; totals sum those retained cells. Most source rolls are from 2017, with Assam and Lakshadweep from 2026. The outputs describe name patterns, not an individual’s residence, origin, or language.
Results follow the appeler inference contract, composition form: every row carries proportions that sum to one, explicit abstention with a machine-readable reason instead of a default distribution, and provenance columns identifying the exact artifacts used.
Installation¶
pip install instate
Usage¶
lookup_state_composition reports the electoral-roll shares for surnames in
the table and abstains on the rest:
import instate
result = instate.lookup_state_composition(["dhingra", "sood", "qzxv"])
result[
[
"surname",
"scored",
"abstention_reason",
"state_share_delhi",
"state_share_punjab",
"surname_record_count",
]
]
# surname scored abstention_reason state_share_delhi state_share_punjab surname_record_count
# dhingra True <NA> 0.530 0.231 7583
# sood True <NA> 0.194 0.364 29451
# qzxv False out-of-dictionary <NA> <NA> <NA>
estimate_state_composition runs the temperature-scaled BiLSTM for the same
quantity, including surnames the table has never seen:
result = instate.estimate_state_composition(["chintalapati"])
estimate_language_composition mixes state evidence with each state’s
Census 2011 mother-tongue shares. By default it uses the lookup where the
surname is known and falls back to the model, recording which in a
language_basis column:
result = instate.estimate_language_composition(["sood", "chintalapati"])
result[["surname", "language_basis", "language_share_punjabi", "language_share_telugu"]]
DataFrame input uses the fleet signature: data first, then the column
name, with every option keyword-only.
import pandas as pd
frame = pd.DataFrame({"lastname": ["sharma", "patel"], "person_id": [1, 2]})
result = instate.lookup_state_composition(frame, "lastname")
Two reference lookups round out the API: lookup_state_official_languages
maps states to their official languages, and list_supported_states returns
the 35-state vocabulary.
Optional coverage adjustment¶
instate.coverage.adjust_surname_counts expands observed surname counts to
explicit, edition-matched electorate totals under a supplied MCAR assumption.
It returns observed counts, estimated counts, weights, and coverage diagnostics
separately. It does not change the lookup, model, or minimum observed-support
threshold. Surname resolution remains in upnaam.
See the coverage guide for a runnable example, output schema, and denominator choices. The working frame is a census-like adult electorate, operationalized by printed roll totals.
What the outputs mean¶
The state shares’ denominator is the surname’s retained occurrences across included rolls. Cells with fewer than three occurrences are excluded before normalization from both lookup and training; their counts do not contribute to the published total. This is not a count of people in the current population. The model targets the same kind of retained-record distribution, with probabilities temperature-scaled against held-out surnames. In 3.3, the lookup incorporates repaired Andaman and Dadra inputs, while the model and its calibration retain their 3.2 training sources. The lookup and model therefore share an estimand, but not an identical source revision.
The language composition is defined, not observed:
p(language | surname) = sum over states of
p(state | surname) x census mother-tongue share of the language in the state
The mother-tongue shares come from Census of India 2011 table C-16, with
Telangana aggregated from its ten 2011 districts and languages below a 1%
share in every state pooled into other
(builder, provenance and
hashes in the shipped manifest). Two caveats are part of the definition:
C-16 records mother tongue, not languages spoken, and the mixing assumes
language and surname are independent within a state, which understates
community-specific associations.
Known data weaknesses: Gujarat surnames remain noisy from OCR;
trailing-vowel spelling variants (Kannada patila,
Odia dasa) are merged into their canonical forms (patil, das).
Abstention¶
A surname the package cannot support gets abstained = True and a reason
from the contract’s shared vocabulary (missing-name, no-letters,
unsupported-script, out-of-dictionary, insufficient-evidence), never a
default distribution. Supported input is romanized ASCII a to z; the
model additionally requires three supported characters.
Model and evaluation¶
Version 3.3 retains the 3.2 checkpoint and its matching calibration. A retrained
candidate had worse record-weighted validation log loss (1.434 versus 1.373),
Brier score (0.263 versus 0.235), and top-three record mass (78.1% versus 79.2%)
on the same 20,000 names with updated targets. Its Karnataka result improved.
The paired surname-bootstrap interval for the log-loss difference includes zero
(-0.016 to 0.131); retaining the existing model is a conservative decision, not
proof of a population-level performance difference. The comparison is recorded
in model_training/roll_recovery_model_diagnostic.json.
The following training details and historical comparison describe the retained 3.2 model, not a model trained on the 3.3 lookup.
The state model is a two-layer character-level bidirectional LSTM trained on 1,483,554 canonical names. Hash assignment fixes the train, validation, and test memberships. A separate hash orders validation names: the first 20,000 choose the epoch with the lowest record-weighted cross-entropy, and the other 165,724 fit one calibration temperature (1.200). Training restored epoch 7 after eight epochs.
These uncalibrated scores use the 20,000 names that chose the checkpoint (4,527,362 retained records). They are development evidence and do not establish generalization. The historical test had already informed development; it was not rescored, and this checkpoint cannot claim an untouched test result.
Checkpoint |
Log loss |
Brier score |
Top-three record mass |
|---|---|---|---|
Released 3.1 |
1.473 |
0.276 |
79.2% |
Earlier initials candidate |
1.553 |
0.288 |
78.4% |
Corrected selection |
1.374 |
0.235 |
79.2% |
Lower log loss and Brier score indicate closer agreement with the retained state distributions. Top-three coverage is the share of record mass assigned to those states. These measures weight source records, not people.
The checkpoint, data, split memberships, and calibration are bound by hashes.
The training and comparison records are in model_training/karnataka_2017_training.json
and model_training/karnataka_2017_model_diagnostic.json. Earlier test results
remain under model_training/history/ and describe earlier checkpoints.
Model artifacts download automatically from a pinned Hugging Face revision.
For offline use, set INSTATE_MODEL_DIR to a directory containing
instate_state_lstm.safetensors, the matching calibration JSON, and lookup Parquet.
Data¶
Sources: parsed electoral rolls and
source PDFs.
Census language shares rebuild from the pinned census downloads with
model_training/build_state_language_shares.py.
Coverage by state¶
The training rolls do not cover every state equally. Coverage below is the table’s record weight divided by the state’s electorate at the 2019 general election; a complete 2017 roll parse sits at 85 to 100 percent. Where a state is short, the model has fewer records to learn its surnames from, and a surname shared with a better-covered state is pulled toward that state.
Coverage |
States |
|---|---|
85 to 100 percent |
Bihar, Odisha, Jharkhand, Goa, Tripura, Manipur, Maharashtra, Meghalaya, Haryana, Chandigarh, Puducherry, Punjab, Madhya Pradesh, Arunachal Pradesh, Mizoram, Uttarakhand, Sikkim, Tamil Nadu, Rajasthan, Uttar Pradesh, West Bengal, Himachal Pradesh |
80 to 85 percent |
Nagaland, Daman and Diu, Telangana (English 2017 rolls, rebuilt in 3.1), Kerala |
55 to 70 percent |
Andhra Pradesh, Delhi |
under 55 percent |
Gujarat (52 percent, OCR loss), Jammu and Kashmir and Ladakh (28 percent, Ladakh and the Jammu region only; the Urdu valley rolls are unparsed) |
2026 roll |
Assam (the 2026 final roll, all 126 constituencies; 113 percent of the 2019 electorate) |
Andaman and Dadra now use edition-matched printed controls rather than the 2019 benchmark above. Parse coverage and usable surname coverage are distinct:
Rebuilt frame |
Printed electorate |
Parsed frame records |
Recorded surname selections |
Retained lookup records |
|---|---|---|---|---|
Andaman, final 2017 |
277,983 |
277,987 |
182,787 |
173,813 |
Dadra, draft 2017 |
217,934 |
217,934 |
155,108 |
147,742 |
Andaman has six residual one-record discrepancies; the source audit retains them rather than deleting or inventing records to force agreement. Dadra uses all 266 draft parts. Final English supplements are missing for 11 parts, so the incomplete final frame is not mixed into draft counts. Relative-only surname candidates remain in the separate resolution artifacts, not observed lookup counts. The retained neural model has not been retrained on these repairs.
Karnataka uses the recovered 2017 archive: 46,549 parts and 40,389,176 active records across 196 of 224 constituencies. The archive omits all 28 Bengaluru constituencies, AC150 through AC177, and 117 parts remain flagged for reconciliation. Upnaam selects a Latin surname for 17,886,612 records (44.3%) using household or relation evidence, or the explicit initials fallback, and abstains on the rest. After the shared filters, lookup and training retain 17,809,983 Karnataka occurrences across 97,381 strings. Elector recovery and surname coverage are different measures; these selective surname counts do not represent the whole Karnataka electorate. Initials fallbacks retain a usable name word without establishing that it is a family surname.
Lakshadweep uses 5,025 Latin surname selections from 57,618 active parsed 2026
entries; the shared lookup/training filters retain 3,312 occurrences across 381
strings. This selective sample has much lower surname coverage than the
complete box parse. The current selection sample has only four
Lakshadweep-bearing names. The lookup supplies direct evidence for covered
names. Regional diagnostics and a
comparison with the previous model are in
model_training/karnataka_2017_model_diagnostic.json. Chhattisgarh is not in
the vocabulary. Per-state sources, build commands, and what each gap would take
are in model_training/prep_er_data/SOURCES.md.
Contributor Code of Conduct¶
The project welcomes contributions from everyone! In fact, it depends on it. To maintain this welcoming atmosphere, and to collaborate in a fun and productive way, we expect contributors to the project to abide by the Contributor Code of Conduct.
License¶
The package is released under the MIT License.
Adjacent repositories¶
appeler/naampy — Infer Sociodemographic Characteristics from Names Using Indian Electoral Rolls
appeler/ethnicolr2 — Ethnicolr implementation with new models in pytorch
appeler/parsernaam — AI name parsing. Predict first or last name using a DL model.
appeler/ethnicolor — Race and Ethnicity based on name using data from census, voter reg. files, etc.
appeler/ethnicolr — Predict Race and Ethnicity Based on the Sequence of Characters in a Name